A fully quantized stochastic coder for low‐bit‐rate speech coding

Abstract

Stochastic coding has the potential of providing high‐quality speech at low bit rates [Schroeder and Atal, ICASSP (1985), pp. 937–940]. In this scheme linear predictive techniques are used to model both the long‐term and short‐term correlations in the speech signal. The remaining residual signal (excitation) is modeled by a codebook populated with samples from a Gaussian source. The whole procedure is based on an analysis‐by‐synthesis approach that minimizes the error between the original and the reconstructed signal. In the coder described by Schroeder and Atal, the excitation was coded at 2 kbps and the remaining parameters were left unquantized. To assess the usefulness of the stochastic coding approach for low‐bit‐rate speech coding, the performance of the coder with all the parameters quantized was investigated. A scalar quantizer was used to encode the short‐term predictor coefficients at 1.8 kbps. Different vector quantization procedures for the long‐term predictor coefficients resulting in bit rates varying from 1 to 2 kbps were investigated. It was found that a total of 8 kbps were required to produce results undistinguishable from the unquantized version. Speech coded at different rates between 4.8 and 8 kbps will be played at the conference. [Work supported by NSA.]